Skip to main content

Training

Status: the training CLI is a stub. soarm-train act and soarm-train diffusion both print their hyperparameters and then exit with "... training not yet implemented." (a TODO(build) in each). The recording and data pipeline described in Dataset Recording & Format is real; the training loops that would consume it are not. Do not report a trained imitation-learning policy from this package. RL training that does run end to end lives in soarm_mjlab — see Reinforcement Learning & Simulation.

Intended CLI shape (once implemented)​

soarm-train {act,diffusion} --repo-id <repo> [--root PATH] \
    --epochs 500 --batch-size 64 --lr 1e-4 --chunk-size 100

Wiring up a real ACT training loop against the existing ChunkedActionDataset/JointNormalizer pipeline is the obvious next step here — the data plumbing needed for it already exists and is tested-by-construction (pure torch/numpy, no hardware).